MicroVMs for agents that snapshot in under 100 ms and fork at will

What it does

AgentENV is a distributed platform that orchestrates large fleets of Firecracker microVMs to host agent environments. It pulls OCI-compatible images on demand using overlaybd, uses local disk as a bounded cache so images can exceed physical capacity, and snapshots both memory and filesystem changes incrementally to S3-compatible storage. It powers the agentic RL training for Kimi K3.

The interesting bit

The platform treats idle environments as cheap: a sandbox pauses in under 100 ms and resumes in under 50 ms, while memory and filesystem snapshots finish incrementally in under 100 ms even under heavy write load. You can also fork a running environment into multiple independent sandboxes, which is a neat trick for parallel agent workflows.

Key highlights

• Boots or resumes snapshot-backed environments in under 50 ms; pauses in under 100 ms.
• Shares the host page cache across storage and memory-snapshot data via ublk for high-performance I/O.
• Supports incremental snapshots to S3-compatible object storage or a shared distributed filesystem.
• Exposes an E2B-compatible HTTP API, so existing E2B SDKs work without code changes.
• Runs on Linux kernel 6.8+ with /dev/kvm access; currently ships without authorization.

Caveats

• The README warns that AgentENV currently does not support authorization, so it should not be exposed to public networks.
• The install script targets Ubuntu 24.04 specifically, though Docker deployment is also documented.

Verdict

Worth a look if you are building agent evaluation pipelines or RL training infrastructure that needs dense, fast-recovering sandboxes. Skip it if you need a multi-tenant public sandbox service out of the box; the missing auth layer means this is strictly for trusted networks today.